Pool topic: Real-time, continuous, and interactive AI video generation. Question revision: 1. Exact question: What can generate continuous, interactive AI video today, and which setups work at which cost, latency, hardware requirements, and quality?
Checked 22 October 2026
FastH3 Omni can finish a 15-second H3 clip in 14 seconds on eight GB200s. That is not a live stream.
Project says: FastVideo released an eight-evaluation, reference-guided H3 model that makes video with synchronized audio. Its 27-workload comparison reports 5.6–11.9× lower time to a complete MP4 than Base H3. The comparison uses eight GB200 GPUs at 1344×768 and 24 fps, with model load excluded.
Reference inputs change the clock
| 15-second request | FastH3 Omni complete MP4 | What it means |
|---|---|---|
| One image | 14.14 seconds | Faster than the 15.08-second output duration on this warm eight-GPU run |
| Nine images and three clips | 59.77 seconds | Almost four times the output duration |
Both figures are project measurements. Base H3 uses a different model and server recipe; matched inputs do not prove equal output quality. A completed MP4 measures neither first frame at a viewer nor response to a new control.
What can you run locally?
The model card lists separate FastVideo recipes for RTX 5090 and DGX Spark using NVFP4 QAD at 480p by default. Its ComfyUI files use their own sampler and component formats. Do not apply the eight-GB200 speed range to a 5090 or a different ComfyUI checkpoint.
Community report: one RTX 3090 user compared four 158-frame, 768×1344/24-fps jobs and averaged 8m26s with Omni versus 8m01s with Turbo+VSA. The author preferred Turbo quality in those samples. That is a small self-report against a different comparator, not a direct contradiction of FastVideo's Base-H3 test.
Next move
If you already have suitable hardware, try one pinned reference-driven clip against your existing workflow. Record cold load, complete-file time, output quality and the actual charged time. If your goal is a controlled live show, wait for or test a receiver path separately: time the first visible frame, a changed prompt, uninterrupted playout and synchronized audio. No such Omni session or full bill was established here.